Co-supervision of a MSc internship at the IUCT-Oncopole in collaboration with IRIT, on a project proposed within my PhD topic and supported by dedicated funding.

Goal: Introduce RLHF-based concepts to improve deformable image registration (DIR) between 3D-MRI and 4D-CT for radiotherapy, by combining deep learning with clinician-guided feedback.

Key points

• Built an ergonomic web application to collect clinician ratings of DIR results
• Defined a practical annotation / scoring strategy for registration quality
• Collected expert feedback and trained an initial reward model
• Proposed a first reinforcement-learning pipeline (RLHF-style) to refine DL-based DIR predictions using human ratings
• Output: dataset + web tool + baseline RL optimization results (conference abstract submission: ESTRO 2026)